// Copyright 2025 The Go Authors. All rights reserved. // Use of this source code is governed by a BSD-style // license that can be found in the LICENSE file. package base import ( "fmt" "os" "runtime" "runtime/debug" "runtime/metrics" "sync" ) // forEachGC calls fn each GC cycle until it returns false. func forEachGC(fn func() bool) { type T [32]byte // large enough to avoid runtime's tiny object allocator var finalizer func(*T) finalizer = func(p *T) { if fn() { runtime.SetFinalizer(p, finalizer) } } finalizer(new(T)) } // AdjustStartingHeap modifies GOGC so that GC should not occur until the heap // grows to the requested size. This is intended but not promised, though it // is true-mostly, depending on when the adjustment occurs and on the // compiler's input and behavior. Once the live heap is approximately half // this size, GOGC is reset to its value when AdjustStartingHeap was called; // subsequent GCs may reduce the heap below the requested size, but this // function does not affect that. // // logHeapTweaks (-d=gcadjust=1) enables logging of GOGC adjustment events. // // The temporarily requested GOGC is derated from what would be the "obvious" // value necessary to hit the starting heap goal because the obvious // (goal/live-1)*100 value seems to grow RSS a little more than it "should" // (compared to GOMEMLIMIT, e.g.) and the assumption is that the GC's control // algorithms are tuned for GOGC near 100, and not tuned for huge values of // GOGC. Different derating factors apply for "lo" and "hi" values of GOGC; // lo is below derateBreak, hi is above derateBreak. The derating factors, // expressed as integer percentages, are derateLoPct and derateHiPct. // 60-75 is an okay value for derateLoPct, 30-65 seems like a good value for // derateHiPct, and 600 seems like a good value for derateBreak. If these // are zero, defaults are used instead. // // NOTE: If you think this code would help startup time in your own // application and you decide to use it, please benchmark first to see if it // actually works for you (it may not: the Go compiler is not typical), and // whatever the outcome, please leave a comment on bug #56546. This code // uses supported interfaces, but depends more than we like on // current+observed behavior of the garbage collector, so if many people need // this feature, we should consider/propose a better way to accomplish it. func AdjustStartingHeap(requestedHeapGoal, derateBreak, derateLoPct, derateHiPct uint64, logHeapTweaks bool) { mp := runtime.GOMAXPROCS(0) const ( SHgoal = "/gc/heap/goal:bytes" SHcount = "/gc/cycles/total:gc-cycles" SHallocs = "/gc/heap/allocs:bytes" SHfrees = "/gc/heap/frees:bytes" ) var sample = []metrics.Sample{{Name: SHgoal}, {Name: SHcount}, {Name: SHallocs}, {Name: SHfrees}} const ( SH_GOAL = 0 SH_COUNT = 1 SH_ALLOCS = 2 SH_FREES = 3 MB = 1_000_000 ) // These particular magic numbers are designed to make the RSS footprint of -d=-gcstart=2000 // resemble that of GOMEMLIMIT=2000MiB GOGC=10000 when building large projects // (e.g. the Go compiler itself, and the microsoft's typescript AST package), // with the further restriction that these magic numbers did a good job of reducing user-cpu // for builds at either gcstart=2000 or gcstart=128. // // The benchmarking to obtain this was (a version of): // // for i in {1..50} ; do // for what in std cmd/compile cmd/fix cmd/go github.com/microsoft/typescript-go/internal/ast ; do // whatbase=`basename ${what}` // for sh in 128 2000 ; do // for br in 500 600 ; do // for shlo in 65 70; do // for shhi in 55 60 ; do // benchcmd -n=2 ${whatbase} go build -a \ // -gcflags=all=-d=gcstart=${sh},gcstartloderate=${shlo},gcstarthiderate=${shhi},gcstartbreak=${br} \ // ${what} | tee -a startheap${sh}_${br}_${shhi}_${shlo}.bench // done // done // done // done // done // done // // benchcmd is "go install github.com/aclements/go-misc/benchcmd@latest" if derateBreak == 0 { derateBreak = 600 } if derateLoPct == 0 { derateLoPct = 70 } if derateHiPct == 0 { derateHiPct = 55 } gogcDerate := func(myGogc uint64) uint64 { if myGogc < derateBreak { return (myGogc * derateLoPct) / 100 } return (myGogc * derateHiPct) / 100 } // Assumptions and observations of Go's garbage collector, as of Go 1.17-1.20: // - the initial heap goal is 4MiB, by fiat. It is possible for Go to start // with a heap as small as 512k, so this may change in the future. // - except for the first heap goal, heap goal is a function of // observed-live at the previous GC and current GOGC. After the first // GC, adjusting GOGC immediately updates GOGC; before the first GC, // adjusting GOGC does not modify goal (but the change takes effect after // the first GC). // - the before/after first GC behavior is not guaranteed anywhere, it's // just behavior, and it's a bad idea to rely on it. // - we don't know exactly when GC will run, even after we adjust GOGC; the // first GC may not have happened yet, may have already happened, or may // be currently in progress, and GCs can start for several reasons. // - forEachGC above will run the provided function at some delay after each // GC's mark phase terminates; finalizers are run after marking as the // spans containing finalizable objects are swept, driven by GC // background activity and allocation demand. // - "live at last GC" is not available through the current metrics // interface. Instead, live is estimated by knowing the adjusted value of // GOGC and the new heap goal following a GC (this requires knowing that // at least one GC has occurred): // estLive = 100 * newGoal / (100 + currentGogc) // this new value of GOGC // newGogc = 100*requestedHeapGoal/estLive - 100 // will result in the desired goal. The logging code checks that the // resulting goal is correct. // There's a small risk that the finalizer will be slow to run after a GC // that expands the goal to a huge value, and that this will lead to // out-of-memory. This doesn't seem to happen; in experiments on a variety // of machines with a variety of extra loads to disrupt scheduling, the // worst overshoot observed was 50% past requestedHeapGoal. metrics.Read(sample) for _, s := range sample { if s.Value.Kind() == metrics.KindBad { // Just return, a slightly slower compilation is a tolerable outcome. if logHeapTweaks { fmt.Fprintf(os.Stderr, "GCAdjust: Regret unexpected KindBad for metric %s\n", s.Name) } return } } // Tinker with GOGC to make the heap grow rapidly at first. currentGoal := sample[SH_GOAL].Value.Uint64() // Believe this will be 4MByte or less, perhaps 512k myGogc := 100 * requestedHeapGoal / currentGoal myGogc = gogcDerate(myGogc) if myGogc <= 125 { return } if logHeapTweaks { sample := append([]metrics.Sample(nil), sample...) // avoid races with GC callback AtExit(func() { metrics.Read(sample) goal := sample[SH_GOAL].Value.Uint64() count := sample[SH_COUNT].Value.Uint64() oldGogc := debug.SetGCPercent(100) if oldGogc == 100 { fmt.Fprintf(os.Stderr, "GCAdjust: AtExit goal %dMB gogc %d count %d maxprocs %d\n", goal/MB, oldGogc, count, mp) } else { inUse := sample[SH_ALLOCS].Value.Uint64() - sample[SH_FREES].Value.Uint64() overPct := 100 * (int(inUse) - int(requestedHeapGoal)) / int(requestedHeapGoal) fmt.Fprintf(os.Stderr, "GCAdjust: AtExit goal %dMB gogc %d count %d maxprocs %d overPct %d\n", goal/MB, oldGogc, count, mp, overPct) } }) } originalGOGC := debug.SetGCPercent(int(myGogc)) // forEachGC finalizers ought not overlap, but they could run in separate threads. // This ought not matter, but just in case it bothers the/a race detector, // use this mutex. var forEachGCLock sync.Mutex adjustFunc := func() bool { forEachGCLock.Lock() defer forEachGCLock.Unlock() metrics.Read(sample) goal := sample[SH_GOAL].Value.Uint64() count := sample[SH_COUNT].Value.Uint64() if goal <= requestedHeapGoal { // Stay the course if logHeapTweaks { fmt.Fprintf(os.Stderr, "GCAdjust: Reuse GOGC adjust, current goal %dMB, count is %d, current gogc %d\n", goal/MB, count, myGogc) } return true } // Believe goal has been adjusted upwards, else it would be less-than-or-equal to requestedHeapGoal calcLive := 100 * goal / (100 + myGogc) if 2*calcLive < requestedHeapGoal { // calcLive can exceed requestedHeapGoal! myGogc = 100*requestedHeapGoal/calcLive - 100 myGogc = gogcDerate(myGogc) if myGogc > 125 { // Not done growing the heap. oldGogc := debug.SetGCPercent(int(myGogc)) if logHeapTweaks { // Check that the new goal looks right inUse := sample[SH_ALLOCS].Value.Uint64() - sample[SH_FREES].Value.Uint64() metrics.Read(sample) newGoal := sample[SH_GOAL].Value.Uint64() pctOff := 100 * (int64(newGoal) - int64(requestedHeapGoal)) / int64(requestedHeapGoal) // Check that the new goal is close to requested. 3% of make.bash fails this test. Why, TBD. if pctOff < 2 { fmt.Fprintf(os.Stderr, "GCAdjust: Retry GOGC adjust, current goal %dMB, count is %d, gogc was %d, is now %d, calcLive %dMB pctOff %d\n", goal/MB, count, oldGogc, myGogc, calcLive/MB, pctOff) } else { // The GC is being annoying and not giving us the goal that we requested, say more to help understand when/why. fmt.Fprintf(os.Stderr, "GCAdjust: Retry GOGC adjust, current goal %dMB, count is %d, gogc was %d, is now %d, calcLive %dMB pctOff %d inUse %dMB\n", goal/MB, count, oldGogc, myGogc, calcLive/MB, pctOff, inUse/MB) } } return true } } // In this case we're done boosting GOGC, set it to its original value and don't set a new finalizer. oldGogc := debug.SetGCPercent(originalGOGC) // inUse helps estimate how late the finalizer ran; at the instant the previous GC ended, // it was (in theory) equal to the previous GC's heap goal. In a growing heap it is // expected to grow to the new heap goal. if logHeapTweaks { inUse := sample[SH_ALLOCS].Value.Uint64() - sample[SH_FREES].Value.Uint64() overPct := 100 * (int(inUse) - int(requestedHeapGoal)) / int(requestedHeapGoal) fmt.Fprintf(os.Stderr, "GCAdjust: Reset GOGC adjust, old goal %dMB, count is %d, gogc was %d, gogc is now %d, calcLive %dMB inUse %dMB overPct %d\n", goal/MB, count, oldGogc, originalGOGC, calcLive/MB, inUse/MB, overPct) } return false } forEachGC(adjustFunc) }